Distributed Optimization for Control and Learning. From Theory to Numerical Software Tools
Organisers
- Giuseppe Notarstefano
- Ivano Notarnicola
- Francesco Farina
- Andrea Camisa
The workshop will run on 11 July 2020 from 10:00 until 17:00 Berlin time
(10am until 5pm CEST/UTC+2h). A recording of the workshop will also be
available for streaming from 12 July until 31 August 2020 for registered
participants.
Practical information for workshop attendees:
The practical sessions are designed for unix-like systems (Linux, MacOS, ...) and will be performed on Ubuntu 18.04. In order to keep up with the live session, please make sure that you have a working installation of MPI (e.g. OpenMPI), the Python 3 interpreter (optionally with virtual environment support), a text editor (e.g. gedit).
For instance, to install OpenMPI in Ubuntu 18.04 systems, run the
following command on a terminal:sudo apt-get install
openmpi-bin libopenmpi-dev.
In MacOS systems, OpenMPI can be installed via homebrew.
Speakers
- Giuseppe Notarstefano, Università di Bologna, Italy
- Ivano Notarnicola, Università di Bologna, Italy
- Andrea Camisa, Università di Bologna, Italy
Summary
Cyber-physical network systems give rise to many important control and learning problems in which solving a constrained optimization problem is a fundamental building block. Optimization problems arising in this context are typically large-scale, (i.e., involve a large set of decision variables and/or constraints). Moreover, in many relevant applications these problems are logically and/or spatially distributed in the sense that the computing units have only partial knowledge of the problem. These features call for a novel computation paradigm, termed distributed optimization, in which agents in a network want to cooperatively obtain an optimal solution to the problem by means of local computation and neighboring communication only. This workshop aims to provide an introductory, theoretical foundation for distributed optimization and a set of advanced challenging problems with selected distributed methods. The methodological part will be supported by the numerical implementation of the presented distributed methods using Disropt, a recently developed Python package for distributed optimization.
Programme
10:00 Welcome and opening remarks
10:15 Introduction to distributed
optimization
10:45 Presentation of
DISROPT package
11:15
Break
11:45 DISROPT: installation and
linear average consensus (practical session)
12:15
Distributed optimization for cost-coupled
optimizaiton
13:00 Lunch
break
13:45 DISROPT: distributed
algorithms for machine learning (practical session)
14:45 Distributed optimization for
constraint-coupled optimization
15:30
DISROPT: distributed primal decomposition for smart grid
control (practical session)
16:30
Concluding remarks
(live discussions will be held between the presentations)